The great physicist Neils Bohr famously joked that “prediction is very difficult, especially about the future.” Bohr had a point, but it is also true that modern predictive modeling techniques are enabling many businesses to innovate, become more efficient, and grow profitably. The predictive algorithms powering the recommendations of Amazon, Netflix, and Google leap to mind. But, as recent books such as Moneyball, Super Crunchers, and The Numerati document, the phenomenon extends well beyond internet and database marketing. Indeed, the transformative power of predictive modeling cuts across such a wide swath of industries that Chris Anderson, the editor of Wired magazine, proclaimed one of today’s most important cultural trends is “the explosion of data about every aspect of our world and the rise of applied math gurus who know how to use it”.
After seeing models built for Property and Casualty insurance evolve from innovative “secret weapons” to table stakes over the past decade, life insurers have begun to investigate whether predictive models can be applied in their industry. Recent experience leads them to answer with a decisive “yes”. Predictive models with a proven ability to help life insurers more effectively target market, make underwriting decisions more economically, accurately, and consistently, enhance retention programs, and even refine product pricing are increasingly creating competitive advantages, which prove to be so difficult to generate in a mature industry.

Understanding the Science – and the Art
At its most basic level, predictive modeling is the process of using known quantities to predict unknown quantities. More precisely, a variety of modeling tools are used to identify linear and non-linear combinations of predictive variables that serve as “leading indicators” of unknown quantities such as the propensity to purchase a product, likelihood of lapse, or mortality. While grounded in proven analytical techniques, a successful predictive modeling project relies just as much on “art” as “science”. For life insurance, this “art” refers to the many interdisciplinary activities needed to bridge the gap between textbook statistics and the business implementation of a successful predictive model. Deep expertise in life insurance, programming, third party data sources, statistical computing, and change management are all essential.

But Will it Work?
Life insurers are constantly looking for operational advantages, but to what extent can easily available data and predictive algorithms preview mortality risk even before receiving an insurance application, or subsequently eliminate the need for expensive and time consuming requirements such as lab tests, medical examinations, and physician’s statements?
The short answer is “the models work quite well.” Until recently, the mortality risk of a life insurance prospect was a mystery. Today, marketing organizations have helped shed light on potential customers nearly all of us. It may be surprising that this marketing data – which includes credit card purchase patterns, self-reported survey results and estimated financial data – would convey information about health, but not when you consider that it might show whether a person regularly purchases running shoes or makes a habit out of eating fast food or has twice as many televisions as we would expect. This data is used to develop disease-predicting submodels which have the ability to proxy such elusive attributes as lifestyle, mortality, and morbidity.
The graph below displays the impressive risk segmentation achieved by a model using only this inexpensive marketing data. A tool with the ability to provide an advance estimate of mortality risk allows insurers to make informed decisions where they traditionally had little insight. For example, a life insurer could score members of a direct marketing database and target risks who fit their marketing strategy (e.g. those most likely to be underwritten as preferred or substandard risks) resulting in more applications generated from people they would prefer to insure, and fewer from those they would not. Alternatively, an insurer could profile the in-force business as underwriting wears off to provide an updated assessment of health status. Armed with better information, insurers could allocate policyholder rewards for wellness programs and target individuals for retention programs. In fact, a life insurer could assess the aggregate health of an entire block of business during M&A due diligence by using a predictive modeling tool and then comparing the results to their own business.
Even more impressive, however, is the segmentation power possible when readily available underwriting data such as application responses and other common database sources are incorporated into the model. The graph above displays the model’s mortality curve against that reflecting the company’s existing underwriting process which, by comparison, included lab results in every case, and often more invasive and costly medical tests.
For many applicants, purchasing insurance is marked by waiting: between application and medical or paramedical requirements, while the lab is processing results, and for the insurer to weigh this information and return a decision. However, this graph shows a model is able to replicate fully-underwritten mortality results for much of applicant pool based only on inexpensive underwriting data available in near real-time. If the rigors of traditional underwriting produce no better results, then why encumber the applicant could score applications immediately upon submission, and jet-issue those that score among the better risks, i.e. where the model and full underwriting produce comparable results, and continue to fully-underwrite those who do not. Perhaps even more significant than the costs savings per case of $150 or more is the improved applicant and agent experience.

What Would Your Mother Say?
The potential benefits of predictive modeling for the life insurance industry are intriguing, but what about legal and ethical questions? Regarding the former, it is important to note that the marketing data described above is not regulated by HIPAA or the Fair Credit Reporting Act, and does not require signature authority. Furthermore, predictive models exclude any data elements that communicate information precluded from traditional life insurance underwriting. What about societal and ethical concerns? Some might suggest that using third-party data is too invasive. Yet the use of such data is accepted in consumer business and property-casualty insurance. Furthermore, good models do not stand or fall on a handful of “silver bullet” variables. Rather a plethora of data elements are combined to “tell the story” that an underwriter would ultimately unveil through more laborious, fallible, and timeintensive methods.
Predictive modeling works, which why it has become ubiquitous in many domains. Still, the life insurance industry has a unique culture and set of norms. One might be concerned that no model could capture the nuance and extensive knowledge required of underwriting professionals. Predictive models hold up extremely well, even on a policy-by-policy basis, to the scrutiny of experienced underwriters. But the models should be regarded as no more than tools that will enable underwriters to fast-track straightforward risks, thus freeing time to evaluate complex risks. Furthermore, expert underwriters can provide insight to the modeling process, thus continuing a virtuous cycle of underwriting excellence that marries human expertise with statistical rigor.
Technological progress forces cultural change and flexible responses have long been central to economic success. While their role will be clarified in the coming years, advanced data mining and predictive analytics will be at the center of life insurers’ search for competitive advantage.